From: Steve G. <sg...@gm...> - 2013-11-30 14:43:03
|
Right, so the challenge becomes your detection process in a verity of scenarios (inside, outside, different lighting conditions, detect all profiles of face, people wearing hats, etc.). The multi-scale pedestrian detection is no simple matter of trying to pick out a shape prior to detection. It has to work in situations where someone is walking behind a bush and only the top half of the person is visible. Even then you still get a lot of false detections with things like tree branches. It also has to work at multi-scale as well and multiple ROIs. The facial recognition I've seen uses 70x70 images to train it. If you are not shrinking the image then you may be wasting CPU if you are feeding it a larger size. Also, I only analyze ROIs that meet a certain thresholds and only when there's motion present. This was illustrated in my example code I linked to. It would increase the detection reliability to run another CV routine like face recognition once a pedestrian has been detected. Offloading this to another process would probably be required since it would not finish before the next frame from the camera is ready. On Fri, Nov 29, 2013 at 9:20 PM, Will Stewart <wil...@gm...>wrote: > A silhouette is only the portion of the full image that actually has > someone moving in it. You can then run a simple face detection, and only > when you detect a face do you actually do a facial recognition. So no > searching all of every frame for a face, and certainly no reducing image > resolution to 320x240. > > As to the question of Java performance, indeed it is no longer the 1990s. > You may want to look to current technical literature on the subject - in > six separate web performance benchmarks<http://www.techempower.com/blog/2013/04/05/frameworks-round-2/>, > Java frameworks took 22 out of the 24 top-four positions. And that's just > the tip of the iceberg; > http://www.infoq.com/articles/9_Fallacies_Java_Performance > > |